Refusal
The model declining a request.
When a model declines to answer or act, usually because of safety training or an explicit instruction. A capability and a failure mode at once.
Over-refusal is a real product problem: legitimate work blocked by a model that reads ambiguity as harm. When you see it, sharpen the context to make the legitimate intent explicit rather than arguing with the model.
Over-refusal is a real product defect, not a safe default: it pushes users toward unsanctioned tools and erodes trust in the system’s judgement. The failure is symmetrical, which is why refusal rates on benign traffic belong in your evaluation alongside the attacks you wanted to stop.
The legitimate request that was blocked
A medical documentation assistant refuses to discuss a drug interaction for a clinician.
The refusal was trained for consumer safety and lands in a professional context where it makes the tool useless. Measure refusal rate on your real distribution, and treat an unnecessary refusal as a bug with the same seriousness as an unwanted answer.
- blogNext-generation Constitutional Classifiers ↗
Classifier-based jailbreak defence and its safety-usability trade-off: the earlier generation reduced jailbreak success but raised refusals on harmless queries.
- paperConstitutional AI: harmlessness from AI feedback ↗
Replaces human harm labels with a written set of principles, a self-critique and revision pass, then reinforcement learning from AI feedback.
Jailbreak
Talking the model past its own restrictions.
Guardrails
Constraints on what the system will do, enforced in code.
Alignment
Getting the system to pursue what we actually intended.
Bias · Fairness
Systematic skew in outputs, usually inherited from data.